Phase 4 Project 4 Image Classification
Presentation Link
A local hospital wants to explore their image recognition options for pneumonia cases. They want eventually replace a few doctors due to staffing issues. They want to test it out for a trial version.
The dataset is from Chest X-Ray of "Normal" and "Pneumonia" images from Kaggle with a total of 5856 .jpeg files. This data set is already broken up into three folders (train, validation and test) and a folder for each category ("NORMAL" or "PNEUMONIA").
Based on the business goal and data, the main metrics will be Recall, F1-Score, and Loss, Accuracy will be secondary.
Iteratively produce models. As new information was learned new models, parameters and transformation techniques were applied.
- Model 1 Baseline
- Model 2-3 Baseline Modification
- Model 4-7 Hyperparameter optimization
- Model 8-10 Transfer Learning
- Final Model (Model 5)
From the 624 Dataset:
- .61% True Positive (383)
- .018% True Negative (112)
- .01% False Negative (7)
- .2% False Positive (122)
Normal:
- Recall:.48
- F1-Score:.63
- Precision:94
Pneumonia:
- Recall:.98
- F1-Score:.86
- Precision:.76
Hospital:
- Important image areas
- Radiologist SME knowledge
Technical:
- Hardware and Software Compatibility (Modeling on M2 GPU Laptop)
- Hyperparameter optimization limits
- Blackbox of Hidden Layers
Hospital:
- Usage: The model is best as a learning tool and not an official diagnosis.
- Strategy: Use the model as an initial reviewer of the images.
- Staffing: The model is best used with a doctor, not standalone.
Technical:
- Scope, review and process images more beforehand
- Visualize the activation functions to see better what areas the model layers are diagnosing
- Iterate model improvement with with augmented data
- Visually inspect the images that were FN and FP.